Classification in Mobility Data Mining

Activity Recognition – Semantic Enrichment

Recognition throughPoints-of-Interest

Given a dataset of GPS tracks of private vehicles, we annotate trajectories with the most probable activities performed by the user.

The method associates the list of possible POIs (with corresponding probabilities) visited by a user moving by car when he stops.

A mapping between POIs categories and Transportation Engineering activities is necessary.

?

Gym LeisureHospital Services

Restaurant Food

The enrichment process

● POI collection: Collected in an automatic way, e.g. from Google Places.

● Association POI – Activity: Each POI is associated to a ``activity". For example Restaurant → Eating/Food, Library → Education, etc.

● Basic elements/characteristics:– C(POI) = {category, opening hour, location}– C(Trajectory) = {duration of the stop, stop location, time of the day}– C(User) = {max walking distance}

● Computation of the probability to visit a POI/ to make an activity: For each POI, the probability of ``being visited'' is a function of the POI, the trajectory and the user features.

● Annotated trajectory: The list of possible activities is then associated to a Stop based on the corresponding probability of visiting POIs

Input & Output

Lat; LonTimeStamp: Sun 10:55 – 12:05

Wd = 500 m

Bank: Mon – Fri [8:00 – 15:30]

Dentist: Mon – Sat [9:00 – 13:00] [15:30 – 18:00]

Church: Mon – Sat [18:00 – 19:00] Sun [11:00 – 12:00]

Primary School: Mon – Sat [8:00 – 13:00]

Input & Output

- Stop: Lat; Lon- TimeStamp: Sun 10:55 – 12:05

Wd = 500 m

Church Services [80%]Bar Food [20%]

Pastry Food [100%]

Inferring Activities from social data

Extraction of personal places from Twitter trajectories in Dublin area

The points of each trajectory taken separately were grouped into spatial clusters of maximal radius 150m. For groups with at least 5 points, convex hulls have been built and spatial buffers of small width (5m) around them.1,461,582 points belong to the clusters (89% of 1,637,346); 24,935 personal places have been extracted.

Examples of extracted places

Statistical distribution of the number of places per person

Recognition of the message topics, generation of topical feature vectors, and summarization by the personal places

Topics have been assigned to 208,391 messages (14.3% of the 1,461,582 points belonging to the personal places)

...

1) Some places did not get topic summaries (about 20% of the places)2) In many places the topics are very much mixed3) The topics are not necessarily representative of the place type

(e.g., topics near a supermarket: family, education, work, cafe, shopping, services, health care, friends, game, private event, food, sweets, coffee)

Obtaining daily time series of place visits and comparison with exemplary temporal profiles

The daily time series of place visits have been obtained through aggregation of daily trajectories using only relevant places for each trajectory. The aggregation was done separately for the work days from Monday to Thursday, and for Saturday, Sunday, and Friday.

Exemplary temporal profiles of different activities or visits to different types of places

The time series of place visits are compared to the exemplary time profiles by means of the Dynamic Time Warping (DTW) distance function. Resulting scores: from 0 (no similarity) to 1 (very high similarity). 15,950 places (64% of all) have no similarity to any of the exemplary time patterns. 4,732 places (19%) have the maximal similarity score of 0.8 or higher; 4,179 of them (16.8% of all) were visited in 6 or more days.

Time series with high similarity to “work” (>=0.8)

1,520 places (6.1% of all). These places have also high similarity to “education”, “transport”, and “lunch”.

The time series similarity scores have been combined with the relative frequencies of the topics using a combination matrix

In 233 places out of the initial 1,520 (15%, 0.9% of all places) the similarity to the “work” profile has been reinforced based on the topic frequencies.

Classification of the places according to the highest combined score (minimum 0.8)

20,247 places (81.2%) are not classified; 4,688 (18.8%) are classified, of them 4,048 (16.2%) were visited in at least 6 days

Activity Recognition – Inductive approach

Eigen-behavioursInput Left: subject’s behavior over the course of 113 days for five

situations / activities Right: same data represented as a binary matrix of 113 days (D)

by 120 (H, which is 24 multiplied by the five possible situations)

Eigen-behavioursMethod Are there key activity distributions from which to infer all others

through linear combination? Same idea as PCA

Key distributions

Eigen-behavioursOutput Set of 3 representative eigen-behaviours Each user's activity can be rewritten as their linear combination

activities

Eigen-behavioursExample

Individual Mobility Networks

How to synthesize Individual Mobility?

Mobility Data Mining methods automatically extract relevant episodes: locations and movements.

Rank individual preferred locations

How to synthesize Individual Mobility?

Graph abstraction based on locations (nodes) and movements (edges)

How to synthesize Individual Mobility?

High level representation

Aggregation of sensitive data

Abstraction from real geography

From raw movement…

… to annotated data

1) Build from data anIndividual Mobility Network Individual Mobility Network (IMN)

3) Use a cascading classification with label propagation (ABC classifier)

2) Extract structural features from the IMN

Extracting the IMN

τ(9) = 8

τ(11) = 3τ(0) = 25

ω(1, 2) = 4

ω(2, 1) = 2

ω(0

, 6) =

4

Extracting the IMN

Trip Features

Length

Duration

Time Interval

Average Speed

Network Features

centrality clustering coefficientaverage path length

predictability entropy

hubbiness degreebetweenness

volume edge weightflow per location

Extracting the IMN

ω(1, 2) = 2

ω(2, 1) = 3

from to weight ccFrom ccTo duration

1 2 2 0.22 0.12 10 min

1 2 2 0.22 0.12 5 min

2 1 3 0.12 0.22 4 min

2 1 3 0.12 0.22 6 min

2 1 3 0.12 0.22 4 min

ABC Classifier● Principles:

– The activities of a user should be predicted as a whole, not separately

– Some activities are easy to classify

– Other activities might benefit from contextual information obtained from previous predictions

● E.g.: a place frequently visited after work might be more likely to be leisure / shopping

ABC Classifier● Inspired by Nested Cascade Classification

ABC Classifier● Inspired by Nested Cascade Classification

The ABC classifier

The ABC classifier

The ABC classifier

The ABC classifier

Experiments

GPS traces

6,953 trips65 vehicles

Semantic Mobility AnalyticsTemporal Analysis

● Pisa traffic

In Out

Semantic Mobility AnalyticsTemporal Analysis

● Calci traffic

In Out

Semantic Mobility AnalyticsTemporal Analysis

User Profiling

In computer science, is the process of construction and extraction of models representing user behavior generated by computerized data analysis.Are employed to study, analyze and understand human behaviors and interactions.Are exploited by many applications to make predictions, to give suggestions etc.

MYWay: Trajectory Prediction

Individual and Collective Profile

Individual ProfileInput: Individual DataOutput: Individual Patterns

Collective ProfileInput: Collectivity DataOutput: Collective Patterns

Prediction using probability mixture models

J. Ghosh, M. J. Beal, H. Q. Ngo, and C. Qiao. On profiling mobility and predicting locations of wireless users. 2006.

J. Ghosh, H. Q. Ngo, and C. Qiao. Mobility profile based routing within intermittently connected mobile ad hoc networks (icman). 2006.

I. F. Akyildiz and W. Wang. The predictive user mobility prole framework for wireless multimedia networks. 2004.

Prediction based on individual and collective preferences

F. Calabrese, G. Di Lorenzo, and C. Ratti. Human mobility prediction based onindividual and collective geographical preferences. 2010.

Prediction using complex networks and probability

D. Barth, S. Bellahsene, and L. Kloul. Mobility prediction using mobile user profiles. 2011.

D. Barth, S. Bellahsene, and L. Kloul. Combining local and global proles for mobility prediction in lte femtocells. 2012.

Collective prediction using t-patterns

F. Giannotti, M. Nanni, F. Pinelli, and D. Pedreschi. Trajectory pattern mining. 2007.

A. Monreale, F. Pinelli, R. Trasarti, and F. Giannotti. Wherenext: a location predictor on trajectory pattern mining. 2009.

Mobility Profiling

A concise model ables to describe the user’s mobility in terms of representative movements, i.e. routines.

This model is called Mobility Profile.

Mining mobility user profiles for car pooling. Trasarti, Pinelli, Nanni, Giannotti. KDD 2011

Derived patterns and models: mobility profiles

User history

An ordered sequence of spatio-

temporal points.

Trips construction

Cutting the user history when a stop is detected

Stops Spatial ThresholdStops Temporal Threshold

Grouping

Performing a density based clustering equipped with a spatio temporal distance

function

Spatial TolleranceTemporal Tollerance

Spatio temporal distance

Profile extraction

The medoid of each group becomes user’s routines and the all

set become the user’s mobility profile

Pruning

Groups with a smallNumber of trips are

Pruned

Support Threshold

Trasarti, Pinelli, Nanni, Giannotti. Mining mobility user profiles for car pooling. ACM SIGKDD 2011

Idea in a nutshell

Use the mobility profile to predict the user’s movements. If it is not able to produce a prediction, a collective predictor is used.The collective predictor is built using the mobility profiles of the crowd.

Experimental setting

Starting from a dataset of 1 month of movements, 5.000 users and 326.000 trajectories. We divided the training set, i.e. 3 weeks and as test set the remaining last week.

The trajectories in the test set are cut to become the queries for the predictor. The cuts tested are taking the first 33% or 66% of the trajectories.

Extracting the Mobility Profiles

The first step is to extract the mobility profiles from the training set. In order to assess the quality of them an empirical analysis is performed.

Routines per user distribution (left), trajectories and routines time start distribution (right) and the dataset coverage (bottom)

Results

MyWay obtains good results which are comparable to a global predictor built on top of the whole set of trajectories.

Test 33% Test 66%

proactive car pooling

Project ICON

Carpooling cycleContext

Several initiatives, especially on the web

Carpooling cycleDistinctive features

Users manually insert and update their rides

Users search and contact candidate pals

Users make individual, “local” choice

System autonomously detect systematic trips

System automatically suggest pairings

System seeks globally optimal allocation

Traditional approach vs. ICON cycle

Carpooling cycleAssumptions

Users provide access to their mobility traces

Carpooling cycleStep 1: Inferring Individual Systematic Mobility

• Extraction of Mobility Profiles– Describes an abstraction in space and time of

the systematic movements of a user.– Exceptional movements are completely

ignored.– Based on trajectory clustering with noise

removalIndividual History Trajectory Clusters Routines

Carpooling cycleStep 2: Build Network of possible carpool matches

Based on “routine containment”− One user can pick up the other along

his trip

Carpooling network− Nodes = users− Edges = pairs of users

with matching routines

passenger

driver

Pro-active suggestions of sharing rides opportunities without the need for the user to explicitly specify the trips of interest.

Matching two routines:

Mobility profile share-ability:

Application: Car pooling

Carpooling cycleStep 3: Optimal allocation of drivers-passengers

• Given a Carpooling Network N, select a subset of edges that minimizes |S|

– S = set of circulating vehicles

provided that the edges are coherent, i.e.:

– indegree(n)=0 OR outdegree(n)=0 (a driver cannot be a passenger)

– indegree(n) ≤ capacity(n)

N

N'

Carpooling cycleThe “simple” ICON Loop

Input mobility data

DM: Extract mobility profiles

Build Carpooling network

CP: Optimal allocation

Users accept/reject suggestions

Carpooling cycleImprovement

• In carpooling (especially if proactive) users might not like the suggested matches– Impossible to know who will accept a given

match– Modeling acceptance might improve results

• Two new components– Learning mechanism to guess success

probability of a carpooling match– Optimization task exploits it to offer

solution with best expected overall success

ML: Learn/update success model

& apply to network

Carpooling cycleRevised ICON Loop

Input mobility data

DM: Extract mobility profiles

Weighted Carpooling network

CP: Allocation with best expected success

Users accept/reject suggestions

Training data

S. Rinzivillo, S. Mainardi, F. Pezzoni, M. Coscia, D. Pedreschi, F. Giannotti

Discovering the Geographical Borders of Human Mobility

KI - Künstliche Intelligenz, 2012.

Networks as a mining tool

Mobility coverages

Step 1: spatial regions

Step 2: evaluate flows among regions

Step 3: forget geography

Step 4: perform community detection

Step 4: perform community detection

Step 5: map back to geography

Step 6: draw borders

Final result

Final result: compare with municipality borders

Borders in different time periodsOnly weekdays movements Only weekend movements

Similar to global clustering: strong influence of systematic movements

Strong fragmentation: the influence of systematic movements (home-work) is missing

Borders at regional scale

Final results

Comparison with “new provinces”